Guitar Tablature Estimation with a Convolutional Neural Network
Andrew Wiggins, Youngmoo E. Kim · Zenodo (CERN European Organization for Nuclear Research) · 2019
Guitar tablature is a popular notation guitarists use to learn and share music. As it stands, most tablatures are created by an experienced guitarist taking the time and effort to annotate a song. As the process is time consuming and requires expertise, we are interested in automating this task. Previous approaches to automatic tablature transcription break the problem into two steps: 1) polyphonic pitch estimation, followed by 2) tablature fingering arrangement. Using a convolutional neural network (CNN) model, we can jointly solve both steps by learning a mapping directly from audio data to tablature. The model can simultaneously leverage physical playability constraints and differences in string timbres implicit in the data to determine the actual fingerings being used by the guitarist. We propose TabCNN, a CNN for estimating guitar tablature from audio of a solo acoustic guitar performance. We train and test our network using microphone recordings from the GuitarSet dataset, and TabCNN outperforms a state-of-the-art multipitch estimation algorithm. We also introduce a set of metrics to evaluate guitar tablature estimation.